Efficient Network Representations Learning: An Edge-Centric Perspective.

Shichao Liu,Shuangfei Zhai,Lida Zhu,Fuxi Zhu,Zhongfei (Mark) Zhang,Wen Zhang

DOI: https://doi.org/10.1007/978-3-030-29563-9_33

2019-01-01

Abstract:Learning succinct/effective representations is of essential importance to network modeling. Network embedding is a popular approach to this end, which maps each vertex of a network to a fixed-length, low-dimensional vector. Previous efforts on network embedding mainly fall into two categories: (1) link based and (2) link + node feature based. In this paper, we provide an edge-centric view where features are associated with edges, in addition to the network structure. The edge-centric view provides a fine-grained characterization of a network, where the dynamic interactions between vertices are considered instead of the static property of vertices. Methodology-wise, we propose an efficient network representations learning approach called the NEEF (network embedding with edge features) model. In particular, the NEEF model seamlessly incorporates the edge features when considering the proximity of a pair of vertices. The model is then trained to maximize the occurrence probability of the neighboring vertices conditioned on the edge features. In our experiments, we show that many of the real-world networks can be abstracted as network with edge features, among which we choose the DBLP coauthor network, the Reddit review network, and the Enron email network for evaluations. Experimental results show that our proposed NEEF model outperforms the other state-of-the-art baselines on tasks such as classification and clustering. We also visualize the learned representations, and show intuitive interpretations of the NEEF model.